La fintech britannique Revolut affirme avoir réduit d'environ 90 % sa facture d'IA de codage en moins d'un an, en remplaçant les outils commerciaux comme Claude Code, l'assistant de programmation d'Anthropic, par un modèle chinois hébergé sur ses propres serveurs.
David Abitbol, qui dirige l'infrastructure de la banque en ligne, a raconté la bascule lors d'une conférence à New York. Ses 2 000 ingénieurs sont passés de Gemini début 2025 à Claude Code début 2026, avant d'adopter GLM, un modèle ouv
La fintech britannique Revolut affirme avoir réduit d'environ 90 % sa facture d'IA de codage en moins d'un an, en remplaçant les outils commerciaux comme Claude Code, l'assistant de programmation d'Anthropic, par un modèle chinois hébergé sur ses propres serveurs.
David Abitbol, qui dirige l'infrastructure de la banque en ligne, a raconté la bascule lors d'une conférence à New York. Ses 2 000 ingénieurs sont passés de Gemini début 2025 à Claude Code début 2026, avant d'adopter GLM, un modèle ouvert développé par le chinois Z.ai et utilisé aujourd'hui en version 5.3.
Le calcul est vite fait. Auto-hébergé, GLM revient à 0,30 dollar le million de tokens, ces unités de texte que les fournisseurs d'IA facturent à la consommation, soit à peu près dix fois moins que Claude Code ou que Codex, son équivalent chez OpenAI. Certains développeurs coûtaient jusqu'à 10 000 dollars par mois, pour un retour jugé décevant en interne.
Le tout tourne sur 56 cartes graphiques Nvidia B200 installées chez Revolut. Plutôt que d'empiler les cartes, les équipes ont surtout gratté de la performance avec des réglages fins, comme la compression du modèle ou la séparation des différentes phases de calcul.
Côté performances, la première réponse tombe en moins de 3 secondes et le système sort environ 200 tokens par seconde, et même 800 le soir quand les bureaux se vident. Au-delà de 5 secondes d'attente, les plaintes pleuvent en interne.
La vraie surprise est ailleurs en fait. Les plus gros gains ne viennent pas de l'écriture de code mais du débogage et du traitement des tickets Jira et Confluence, ces outils de suivi de projet que tout le monde adore détester. Le temps de codage se serait quand même amélioré de 60 %.
Et puis il y a l'effet psychologique. Avec des tokens illimités en interne, les ingénieurs ne rationnent plus leurs requêtes pour tenir le budget du trimestre, et la facture ne réserve plus de mauvaise surprise en fin de mois. Les modèles commerciaux restent accessibles pour les cas particuliers, les tâches visuelles partent par exemple chez Gemini.
Un modèle chinois au coeur d'une banque, la question fâche forcément. Abitbol balaie. Hébergés en local ou sur un cloud régional, les poids du modèle ne font selon lui courir aucun risque, le danger serait plutôt de passer par des API hébergées on ne sait où.
Les fenêtres d'un million de tokens que vendent les gros acteurs du secteur sont à ses yeux "une arnaque", 256 000 suffisent largement à coder. Tout le marché vend exactement l'inverse depuis deux ans.
Google a annoncé mardi que Gemini, Gemini Notebook et même son moteur de recherche seront désactivés par défaut sur les comptes des élèves américains, de la maternelle au lycée, tant que les administrateurs scolaires ne les auront pas activés eux-mêmes. Les nouveaux réglages arriveront dans les prochains mois, pour une bascule prévue début 2027.
Présenté comme ça, le geste ressemble à une concession faite aux enseignants. Le détail qui fâche est sorti dans le New York Times, où Chris Phillips, l
Google a annoncé mardi que Gemini, Gemini Notebook et même son moteur de recherche seront désactivés par défaut sur les comptes des élèves américains, de la maternelle au lycée, tant que les administrateurs scolaires ne les auront pas activés eux-mêmes. Les nouveaux réglages arriveront dans les prochains mois, pour une bascule prévue début 2027.
Présenté comme ça, le geste ressemble à une concession faite aux enseignants. Le détail qui fâche est sorti dans le New York Times, où Chris Phillips, le responsable de l'éducation chez Google, a confirmé qu'une école qui refuse l'IA perdra du coup l'accès à Search, parce que la recherche et l'IA sont maintenant profondément intégrées, selon ses propres mots. Les élèves verront simplement un message leur expliquant que le service n'est pas disponible.
Devant les critiques, Google a publié le soir même une mise au point sur X pour assurer que l'annonce donne aux écoles plus de choix, pas moins, et que chaque service peut être activé individuellement, un, deux ou les trois.
Sauf que cette réponse esquive le vrai sujet. Activer la recherche revient à accepter les résumés d'IA affichés en haut des résultats, puisque Google ne propose aucune version du moteur qui en serait débarrassée. La configuration que réclament beaucoup de profs, garder la recherche pour apprendre aux élèves à trouver des sources tout en bloquant les réponses toutes faites, n'existe tout simplement pas au catalogue.
Google sait pourtant ce que coûte la méthode forte. En août, des fonctions d'IA activées d'office dans Classroom avaient mis les enseignants américains en colère, et la firme avait fait machine arrière.
Les Chromebooks et les outils Google équipent une part énorme des écoles américaines, et un district scolaire ne remplace pas son moteur de recherche du jour au lendemain.
Pour les établissements qui diront oui, Google met en avant Guided Learning, un tuteur qui décompose les exercices étape par étape au lieu de donner la réponse, histoire de limiter la triche. Les enseignants pourront même consulter les transcriptions des conversations entre les élèves et le chatbot.
Le calendrier est quand même osé. New York et Los Angeles ont voté en septembre un moratoire d'un an sur l'IA générative dans leurs écoles publiques, et l'État de New York interdit depuis la semaine dernière certaines fonctions de chatbot aux mineurs.
Malaysian SMEs could test AWS's Quick and Kiro for six months without subscription fees, but eligibility rules and post-trial pricing remain unclear.
The post AWS Wants 10,000 Malaysian SMEs Using AI by 2028 — Offers Six Months of Free AI Tools appeared first on TechRepublic.
Singapore, Malaysia and Thailand are preparing for Xiaomi's new flagships, but launch plans and hardware details vary across Southeast Asia.
The post Xiaomi 18 Pro Heads to Southeast Asia — Which Countries Have Confirmed Launch Plans? appeared first on TechRepublic.
Samedi, pendant que vous étiez tous en manif, Andrej Karpathy a demandé à Claude de lui fabriquer un YouTube à partir de toutes les merveilleuses vidéos que les gens se recommandent sur HackerNews et ça
a donné naissance
à
HackerTube
. Déjà rien que ça c'est top !
Mais moi ce qui m'a encore plus c'est que le lendemain Roansh a fait pareil avec les bouquins en laissant Claude coder
HackerBooks
.
Et voilà comment
Samedi, pendant que vous étiez tous en manif, Andrej Karpathy a demandé à Claude de lui fabriquer un YouTube à partir de toutes les merveilleuses vidéos que les gens se recommandent sur HackerNews et ça
a donné naissance
à
HackerTube
. Déjà rien que ça c'est top !
Mais moi ce qui m'a encore plus c'est que le lendemain Roansh a fait pareil avec les bouquins en laissant Claude coder
HackerBooks
.
Et voilà comment on se retrouve avec une bibliothèque de 15 766 livres tirés de 281 363 mentions HN publiées depuis 2007. Et un livre n'y figure que s'il a été cité par au moins 2 personnes différentes. Ensuite chaque fiche renvoie aux commentaires qui en parlent et il y a même des suggestions à la "For You" qui vous fait des propositions raccord avec vos goûts selon ce que vous avez ouvert comme fiche.
Ensuite, l'IA a fait un tri pour distinguer les citations des recommandations. Par exemple The Art of Computer Programming de Knuth cité par 1 125 personnes mais recommandé par 336, alors que How to Win Friends and Influence People de Dale Carnegie en compte 579.
Y'a même Mein Kampf, le livre préféré de ceux qui ont été bercés trop prêt du mur, qui a été cité (je pense pour la blague) par 535 personnes mais a été fièrement recommandé quand même par 15 personnes.
Bien sûr, on ne peut pas télécharger les livres gratuitement dessus. Ça renvoie vers Amazon et compagnie. Mais ça permet de prendre un petit peu la température sur les classiques, organisés dans différentes rubriques comme fiction, programmation, mathématiques, philosophie, fantaisie, etc.
Et si votre livre préféré manque à l'appel, ou si Claude l'a confondu avec un autre, n'hésitez pas à le signaler à roansh
Spotify Jam lets party guests add songs from their own phones. Here's how to set up a shared queue, invite friends, and manage playback controls.
The post Hosting a Halloween or Holiday Gathering? Here’s How to Use Spotify Jam appeared first on TechRepublic.
Vietnam and India are advancing AI, cybersecurity, and digital payments partnerships, but key projects and cross-border payment services remain in development.
The post Vietnam and India Push AI, Cybersecurity, and Cross-Border Payments at Tech Forum appeared first on TechRepublic.
Vietnam and India are advancing AI, cybersecurity, and digital payments partnerships, but key projects and cross-border payment services remain in development.
Greece’s under-15 social media ban starts Jan. 1, 2027. Learn who is affected, how age checks will work, and what users should know about privacy.
The post Greece Passes Under-15 Social Media Ban Starting January appeared first on TechRepublic.
Meta's Ray-Ban and Oakley AI glasses are now available in Malaysia. Compare four models, their prices, features, and prescription compatibility.
The post Meta AI Glasses Now on Sale in Malaysia: Which Ray-Ban or Oakley Model Should You Buy? appeared first on TechRepublic.
Nvidia's latest RTX driver adds shader stutter improvements. Here's which GPUs and games support the feature and what it means for PC performance.
The post Nvidia’s New Driver Targets Game Stutter — Which RTX GPUs and Games Benefit? appeared first on TechRepublic.
Nvidia and Agile Robots unveiled Agile ONE SN, a compact humanoid with Jetson Thor computing and Isaac GR00T training tools. Here's what it can do, its specifications, and when it launches.
The post Agile ONE SN Humanoid Robot Bundles Nvidia’s Robotics Stack appeared first on TechRepublic.
Nvidia and Agile Robots unveiled Agile ONE SN, a compact humanoid with Jetson Thor computing and Isaac GR00T training tools. Here's what it can do, its specifications, and when it launches.
Nvidia's 64GB DGX Spark launches at $4,999 with the same GB10 chip as the 128GB version. Compare memory limits, model support, clustering costs, and what IT buyers should verify.
The post Nvidia DGX Spark 64GB Launches at $4,999: Is It Worth It? appeared first on TechRepublic.
Nvidia's 64GB DGX Spark launches at $4,999 with the same GB10 chip as the 128GB version. Compare memory limits, model support, clustering costs, and what IT buyers should verify.
Nvidia and Hugging Face show how a $399 duck-shaped robot learns to walk using AI simulation. Here's how Microduck works, what it costs, and when buyers may get one.
The post Nvidia Teaches a $399 Duck Robot to Walk Using AI appeared first on TechRepublic.
Nvidia and Hugging Face show how a $399 duck-shaped robot learns to walk using AI simulation. Here's how Microduck works, what it costs, and when buyers may get one.
Australia is preparing legal action against Meta, TikTok, Snap, and Google over alleged breaches of its under-16 social media ban. Here's what could happen next.
The post Australia Prepares to Sue Meta, TikTok Over Under-16 Social Media Ban appeared first on TechRepublic.
Australia is preparing legal action against Meta, TikTok, Snap, and Google over alleged breaches of its under-16 social media ban. Here's what could happen next.
Explore the Microsoft Surface Laptop Ultra's price, specs, Nvidia RTX Spark AI performance, battery life, and key features before buying.
The post Microsoft Surface Laptop Ultra Cheat Sheet: Price, Specs, Features, and More appeared first on TechRepublic.
Nvidia RTX Spark laptops start at $2,599. Explore local AI capabilities, compatibility checks, and the cost and security trade-offs IT teams should assess.
The post Nvidia RTX Spark Brings Local AI Agents to Windows Laptops appeared first on TechRepublic.
Nvidia RTX Spark laptops start at $2,599. Explore local AI capabilities, compatibility checks, and the cost and security trade-offs IT teams should assess.
The RTX 5080 offers more processing power, but the RTX 5070 Ti shares its 16GB VRAM and DLSS capabilities. Here's how performance, power, and 2026 pricing compare.
The post RTX 5070 Ti vs RTX 5080: Which GPU Is Better Value in 2026? appeared first on TechRepublic.
The RTX 5080 offers more processing power, but the RTX 5070 Ti shares its 16GB VRAM and DLSS capabilities. Here's how performance, power, and 2026 pricing compare.
RTX Spark is NVIDIA's new local AI-focused Windows PC platform, combining Blackwell RTX graphics and Grace CPU technology into a single processor that brings secure LLMs, high-end creative workflows, and gaming performance to laptops and desktops. With that, NVIDIA positions RTX Spark as a new generation of Windows PCs, with Surface Laptop Ultra as the flagship device.OEMs are also on board, with RTX Spark laptops launching alongside Microsoft's Surface from bran
RTX Spark is NVIDIA's new local AI-focused Windows PC platform, combining Blackwell RTX graphics and Grace CPU technology into a single processor that brings secure LLMs, high-end creative workflows, and gaming performance to laptops and desktops. With that, NVIDIA positions RTX Spark as a new generation of Windows PCs, with Surface Laptop Ultra as the flagship device.
OEMs are also on board, with RTX Spark laptops launching alongside Microsoft's Surface from brands like HP, Dell, ASUS, and MSI. Our team at Windows Central has followed NVIDIA's efforts closely since its inception and has since gone hands-on with its wave of launch devices. This is everything we know so far about RTX Spark and what we expect to see in the future.
The first wave of RTX Spark laptops just started shipping and hit store shelves. Microsoft's Surface Laptop Ultra likely turned the most heads, but there are actually seven different laptops powered by NVIDIA's new superchip (more if you count different sizes of similar laptops as separate devices).ASUS, Dell, HP, Lenovo, Microsoft, and MSI all have new laptops with the RTX Spark inside, and reviews for many of those PCs just went live. We're still waiting for re
The first wave of RTX Spark laptops just started shipping and hit store shelves. Microsoft's Surface Laptop Ultra likely turned the most heads, but there are actually seven different laptops powered by NVIDIA's new superchip (more if you count different sizes of similar laptops as separate devices).
ASUS, Dell, HP, Lenovo, Microsoft, and MSI all have new laptops with the RTX Spark inside, and reviews for many of those PCs just went live. We're still waiting for reviews of several RTX Spark laptops, but first impressions are in.
Reviews for these devices rolled out a bit differently than other product launches. Experts and reviewers were allowed to share their thoughts about the RTX Spark a day before publishing any reviews of laptops with the RTX Spark inside.
So, if you want to see what people have to say about the processor, check out our RTX Spark review roundup. Below is a collection of opinions about the first wave of RTX Spark laptops.
Note that these reviews were made possible thanks to review samples provided by NVIDIA or laptop makers.
Surface Laptop Ultra: Reviews
Windows Central (4/5): "It's a laptop for those who know they need it, and it's mainly for AI developers and high-end content creators. If you find yourself leaning on local agents to complete tasks for you, or enjoy generating images and videos using AI models, this is the laptop for you."
Tom's Hardware (3.5/5):"If you want the nicest Surface Laptop Microsoft has ever released, the Surface Laptop Ultra is for you. The system definitely copies the MacBook Pro in some areas, but is a luxury machine with a solid build, exceptionally bright display, a clever charging connector, and powerful components. It's the most exciting Surface hardware in years, and I hope it’s a signal that Microsoft is looking at the brand for more reinvention."
The Verge (6/10):"Microsoft has come closer than ever to building a Windows alternative to the MacBook Pro. The hardware looks and feels great, and Nvidia’s GPU brings CUDA support and decent gaming capabilities to a high-end laptop design. But the Surface Laptop Ultra’s biggest selling point is also its biggest question mark: local AI. It all feels like a work in progress."
Wired (7/10):"The Surface Laptop Ultra is a true MacBook Pro competitor for Windows—if only it weren't for the lack of battery life."
Gizmodo: "It’s not the most performant laptop available today. It’s not the cleanest PC experience, either. It is, however, a laptop that can finally close the gap between performance and battery life. With all the wrinkles ironed out, the Surface Laptop Ultra is the laptop I wish I could carry with me for work, pleasure, or whatever the AI overlords think to do with my degrading body next."
Asus ProArt P16: Reviews
Tom's Guide (4/5):"The Asus ProArt P16 is an outstanding laptop for creativity, with its legacy continuing with the might of Nvidia's RTX Spark N1X. With an even slimmer design that's light enough to bring with you anywhere, a knockout OLED touchscreen with amazing brightness and colors, the right ports and battery life that can (finally) get you past a workday and still have juice left, Asus has done its creator laptop justice."
Trusted Reviews (4.5/5): "The Asus ProArt P16 (2026, RTX Spark) is a slim, stunning creator laptop that shows what RTX Spark can do, with standout integrated graphics, capable performance and nearly 14 hours of battery life. Gaming is a little behind a discrete RTX 5070 and the screen is slightly dim next to other ProArts, but it’s an impressive debut for Nvidia’s Windows chip."
Notebookcheck (9/10):"The new ProArt P16 is a stunning laptop and even after using it for more than a week, we are still baffled how thin and light this 16-inch laptop really is without compromising in terms of stability, build quality or features. "
Wccftech: "The ASUS ProArt P16 with NVIDIA RTX Spark N1X feels like a genuine first-class Windows PC rather than a compromised “Windows on Arm” experiment."
HP OmniBook Ultra 16: Reviews
Tom's Guide (4/5): "The HP OmniBook Ultra 16 isn’t a laptop for general customers, but it’s great at what it was designed for. If you want a portable device for on-device AI work or creative projects, you won’t go wrong here. The fact that the laptop is a beautiful piece of work only sweetens the deal."
RTX Spark laptops reviews
Overall, reviews of the first RTX Spark laptops are positive. Unsurprisingly, they focus largely on the RTX Spark itself. That makes sense considering many of these laptops have sibling devices with chips from other companies.
Many called the RTX Spark laptops "MacBook Pro competitors," which is praise when coming from just about any laptop reviewer. Being mentioned alongside the MacBook Pro remains one of the highest compliments a laptop can receive.
Like the RTX Spark itself, these laptops should be viewed in the context of their target audience. They're primarily designed for creators and developers, though different models lean more heavily toward one group or the other.
RTX Spark laptops like the Surface Laptop Ultra earned praise from reviewers for GPU performance. (Image credit: Zac Bowden / Windows Central)
While they're capable of playing games, gaming is not the primary focus of these devices. None of them are gaming laptops at this point, even though they're capable of playing many games.
But with so many devices launching today, we're still waiting for hands-on experience with quite a few RTX Spark laptops. Publications have taken different approaches to covering these laptops since so many RTX Spark devices launched at once. Some got their hands on as many as possible and are working to review each one separately. Others focused on the Surface Laptop Ultra, presumably due to availability and the fact that it's made by Microsoft.
Built around the NVIDIA RTX Spark platform with an N1X CPU, RTX GPU, and up to 128GB of unified memory, the Surface Laptop Ultra hits new performance heights for power users. It's all contained in a gorgeous chassis with a mini-LED display and tons of ports. Prices start at $2,599.99, and units will begin shipping on October 16, 2026.
The new ASUS ProArt P14 and P16 duo gives users more size options while retaining the same great RTX Spark performance. The P14 is the lightest ProArt laptop ever, while the P16 is the way to go for a larger OLED display. Prices start at $2,599.99 for the P14 and $2,799.99 for the P16. They're expected to begin shipping on October 16, 2026.
Dell's new XPS 16 with RTX Spark features a 3.2K OLED display, lattice-free keyboard, and the usual premium XPS styling. Prices start at $3,799.99 and units are expected to begin shipping on October 16, 2026.
The OmniBook X 14 is the smaller of the two new PCs, but it features the same RTX Spark options for maximum performance. Models start at $2,999.99 and are expected to ship beginning November 4, 2026.
The larger of the two new HP RTX Spark laptops delivers a gorgeous 16-inch OLED display with 3K resolution and, of course, your choice of RTX Spark chips. Prices start at $3,199.99, and shipments are expected to begin on November 4, 2026.
Lenovo's Yoga 9n should be the right choice for those who value a versatile convertible design and premium build quality. Prices start at $3,099.99, and models are expected to begin shipping on October 16, 2026.
MSI's Prestige N16 Flip AI+ is another versatile convertible featuring the RTX Spark platform and a Tandem OLED display. It appears that preorders might kick off on October 8 for this laptop, with prices starting at $3,299.99.
Microsoft Maia 300 and Nvidia Rubin target different AI infrastructure needs. Compare their memory, bandwidth, efficiency claims, availability, and deployment options.
The post Microsoft Maia 300 vs Nvidia Rubin: Specs, Not Benchmarks appeared first on TechRepublic.
Microsoft Maia 300 and Nvidia Rubin target different AI infrastructure needs. Compare their memory, bandwidth, efficiency claims, availability, and deployment options.
Connexion
